Replace the Harness. Preserve the Session.

The Architectural Observation

Long-running agents introduce a familiar distributed-systems problem.

The process carrying execution can fail while the underlying work remains valuable.

A model invocation may complete. A tool may modify an external system. An agent may accumulate context over several hours.

The execution environment can then become unavailable.

The architectural requirement is continuity of the work across that interruption.

Anthropic's Managed Agents engineering architecture provides a concrete example.

Its design separates three responsibilities.

The session records the history of agent activity.

The harness manages model interaction and tool coordination.

The sandbox provides an environment for code execution and file operations.

These components operate through defined interfaces.

Their implementations can evolve independently.

The Importance of State Ownership

A harness is an execution mechanism.

Its implementation may contain model-specific assumptions and orchestration behavior.

Those assumptions can change as model capabilities improve.

Anthropic found that its earlier context-reset workaround had become unnecessary after observing different behavior in a newer model.

This illustrates the temporary nature of some harness mechanisms.

The session record serves a different purpose.

It provides a durable account of work across changes in the components carrying execution.

The distinction creates an architectural advantage.

A replacement harness can reconstruct relevant execution context from the session history.

A replacement sandbox can be provisioned according to the workflow's requirements.

The enterprise can preserve the continuity of work while allowing implementation components to evolve.

Session History and Business State

Session history is one category of durable information.

Workflow state represents another.

Business-system state establishes the authoritative effects of external operations.

These categories require distinct ownership.

A session may record that an agent requested a payment.

A workflow checkpoint may record that the payment operation remains pending.

The payment system establishes whether funds were transferred.

The architecture must connect these records through consistent execution and transaction identities.

This relationship matters during recovery.

A replacement worker requires sufficient information to determine the next valid action.

The authoritative external system supplies evidence about completed effects.

Established Engineering Foundations

Durable workflow systems provide relevant engineering foundations.

Temporal records workflow events and uses deterministic replay to reconstruct execution progress.

LangGraph provides checkpoint-based persistence for graph state. Its stores support application-defined information across threads.

These mechanisms serve different execution models.

They demonstrate the importance of persistent state and explicit recovery boundaries.

Their presence establishes useful infrastructure capabilities.

The complete business workflow must still account for authorization and external transaction semantics.

Recovery as an Architectural Transition

Consider an agent that prepares a customer refund.

The workflow records the approved refund amount and initiates the transaction.

The execution sandbox terminates while the payment service processes the request.

A new sandbox becomes available.

The replacement harness retrieves the session history and workflow checkpoint.

The payment service provides the authoritative transaction status.

The workflow uses that evidence to determine the appropriate continuation.

The replacement process operates under current authority and the original transaction identity.

The execution environment changes.

The business obligation remains traceable.

This is an illustrative architecture scenario.

Design Requirements

A durable agent execution architecture should establish five responsibilities.

Persistent execution identity

The workflow retains an identity across process and infrastructure changes.

Durable event history

Consequential execution events remain available according to the required retention policy.

Explicit state ownership

Session history and workflow state have defined custodians. External systems remain authoritative for their own business effects.

Recovery qualification

A replacement worker reconstructs relevant state and reconciles uncertain operations before consequential continuation.

Current authorization

Recovered execution evaluates applicable permissions at the next action boundary.

These responsibilities support continuity across changing harness implementations.

The Architectural Principle

A harness carries the work.

A durable session preserves its execution history.

A workflow engine coordinates continuation.

Business systems establish the effects of completed actions.

The enterprise gains resilience when the lifetime of its obligations extends beyond the lifetime of its execution components.

Research Foundations

Anthropic. Scaling Managed Agents: Decoupling the Brain from the Hands. April 2026.

First-party engineering account of separate session, harness, and sandbox interfaces.

https://www.anthropic.com/research/managed-agents

Temporal. Workflow Execution Overview.

Technical documentation covering durable workflow execution and event-history replay.

https://docs.temporal.io/workflow-execution

LangGraph. Persistence.

Technical documentation describing checkpoints and application-level stores.

https://docs.langchain.com/oss/python/langgraph/persistence

Connected Work

The Enterprise Agent Control Plane Flagship research publication.

Execution Continuity Framework Original framework for transition-specific verification.